# Extensions of Averaging Theory Power Electronic Systems

Ηλεκτρονική - Συσκευές

24 Νοε 2013 (πριν από 4 χρόνια και 5 μήνες)

101 εμφανίσεις

542
IEEE TRANSACTIONS ON POWER ELECTRONICS,
VOL
11,
NO
4,
JULY
1996
Extensions of Averaging Theory
for
Power Electronic Systems
Lehman and Richard
M.
Bass,
Senior Member, IEEE
Abstruct-
This paper extends averaging theory for power
electronic systems to include feedback Controlled converters. New
averaging techniques based on the integral equation description
provide theoretical justification for commonly used averaging
methods. The new theory provides a basis for answering fun-
damental questions about the averaging approximation.
A
ripple
estimate expression is presented, along with the simulation results
for a feedback controlled boost converter.
I.
INTRODUCTION
TATE space averaging techniques are commonly used in
the analysis and control design
of
pulse width modulated
(PWM) power electronic systems [I]-[3]. However, it was
not until recently that rigorous mathematical justification [3],
[4] was given that theoretically explained the applications
of
these averaging techniques.
As
[3] and
[5]
have pointed out,
the theoretical development of PWM systems lags far behind
the many practical control applications.
In
[ 3],
classical Russian averaging techniques
[6],
[7]
are
shown to be applicable to several types of
PWM
power elec-
tronic systems, such as open loop dc-dc converters. Besides
using these classical averaging techniques to prove stability.
[31 also gives
a
ripple estimate for improving the accuracy
of
the averaging technique, even for systems with large ripple.
However, the application of the results of [3] is limited to
systems with time discontinuities.’
In fact, the classical averaging theory used in [3] is not
applicable when there are state discontinuities. This is sig-
nificant because all feedback controlled converters are state
discontinuous. In
[3],
the argument is made that smooth
commutation models can be used in place of the discontinuous
Heaviside unit step function to avoid any state discontinuity
in the mathematical system model. In essence, this idea
was introduced by Filoppov
[SI
to justify what is meant by
solutions to state discontinuous differential equations. The
work of [9] continues this line of thinking by presenting
Manuscript received September 10, 1993; revised December
7,
1995.
B. Lehman was supported by the National Science Foundation under a
Presidential Faculty Fellowship, CMS-959268.
B.
Lehman is with the Department
of
Electrical and Computer Engineering,
Northeastern University, Boston, MA 021 15 USA.
R.
M. Bass is with the School
of
Electrical and Computer Engineering,
Georgia Institute of Technology, Atlanta, GA 30332-0250 USA.
Publisher Item Identifier
S
0885-8993(96)03562-4.
‘In this paper, a system with “time discontinuity” is described by
a
differential equation whose right-hand side is discontinuous with respect
to time. A system with “state discontinuity” is described by a differentia1
equation whose right-hand side is discontinuous with respect to a state
variable.
stability results which rely
on
abstract averaging theory (see
references in [9]) that partially combine the results of [6] and
171
with the theory of Filoppov [SI.
It is the purpose of this paper to introduce averaging
techniques that are general enough to encompass both time
discontinuity and large classes of state discontinuity, without
utilizing the (difficult) theory of Filippov. Because the proofs
are straightforward (essentially relying on the Fundamental
Theorem
of
Calculus and Gronwall’s inequality), insight on
both transient and asymptotic behavior of
PWM
feedback
controlled dc-dc converters is obtained. The results of this
paper begin to provide theoretical justification for commonly
used averaging techniques. In addition, this work points out
some shortcomings in the averaging technique (which to
our knowledge have not been documented before). Some
readers may question whether there is a significant contribution
in writing a paper that theoretically justifies models that
have been in use for
so
many years. However, we believe
that it is vital to bridge theory with practice in order for
future fundamental contributions to be made. In fact, the
theoretical results of this paper have led to the discoveries
of
new, more accurate switching-frequency-dependent-averaged
models [ I O], published in
a
separate paper.
Section I1 reviews some
of
the mathematical issues asso-
ciated with state discontinuous systems. The primary theoret-
ical contribution of this paper is contained in two theorems
presented in Section 111. Section IV discusses the practical
implications of the results of Section
111
and gives numer-
ical examples and computer simulations. Section
V
draws
conclusions.
11.
THEORETICAL PRELIMINARIES
The difficulty in mathematically justifying averaging
approximation techniques
of
state discontinuous differential
equations can be best explained through an example. Consider
the state discontinuous differential equation
k( t )
=
f(z)
+
bu(d(z)
-
tri
(t:
7’))
(2.1)
where
x
E
R”,
b
E
R”,
f:
R”
i
R”
and
d:
72”
i
R
are
both continuous functions with
0
5
d(x)
5
1,
and
U(.)
is the
Heaviside step function, i.e.,
u( s)
=
1
for
s
2
0 and
u( s)
=
0
for
s
<
0. The function tri
(t:
T )
=
( t/T)
-
f l oor( t/T)
=
(t
mod)
TIT
is shown in Fig.
1.
Equation
(2.1)
is
a
typical
representation
of
a
feedback controlled
PWM
Buck converter
[3]. The theory presented in [3], however, only applies to
0885-8993/96\$05.00
0
1996 IEEE
LEHMAN AND
BASS:
EXTENSIONS OF AVERAGING THEORY
FOR POWER ELECTRONlC
SYSTEMS 543
t r i ( t,T )
Fig.
1.
Tri(t,
T )
open-loop control and does not extend to feedback controlled
converters.
The usual condition for a unique solution of (2.1) to exist
is that the right-hand side satisfy a Lipschitz condition.
(A
function,
f ( z )
is said to be Lipschitz with constant
k
>
0
if
I l f (z)
-
f(y)II
5
k/l z
-
y/I
for any
z
E
R",
y
E
R".)
However (2.1) is not Lipschitz since it is discontinuous with
respect to
2.
Hence, standard approaches fail when trying to
prove the existence of a unique solution-which implies that
formal averaging approximations of (2. 1) cannot, in general,
be directly derived. There is an extensive amount of literature
on differential inclusions that shows how one can redefine
what is meant by a unique solution to (2.1) (see Filippov [SI).
However, this paper shows that, under the standard operating
conditions of power electronic systems (no chattering), the
theory of differential inclusions is not needed to theoretically
justify averaging approximations.
While in general "standard" solutions to (2.1) are not known
to exist, under the proper conditions (see Section 11-A), there
are a finite number of jumps in the right-hand side of
(2.1)
on any finite time interval, and each jump (switch) is norm
bounded due to the fact that 0
5
U(.)
5
1.
This implies
that (under these conditions) the right-hand side of (2.1) is
Lebesgue integrable for all
t
2
t o
and that the solution of the
integral equation
z(t;
t o,
z(to))
=z ( t )
E
4 t o )
+
1;
[ f ( z ( s ) )
+
bu( d( z( s) )
-
tri
(s,
T) ) ]
ds
(2.2)
is unique and satisfies state differential equation
(2.1)
almost
everywhere. Hence, when no chattering occurs in the system,
the "standard" solution to (2.1) can be derived and will
be equal to the solution
of
integral equation (2.2) almost
everywhere.
Furthermore, when there is no chattering,
z(t;
t o,
.(to))
=
x( t ),
as given by (2.2), is a continuous function that depends
continuously on its switching period,
T.
Using this fact, [4]
develops approximation techniques by examining (2.2) instead
of
(2.1). This work by Sira-Ramirez shows that the solution
of (2.2) can be accurately approximated by an autonomous
averaged system by letting
T
-+
0.
In [4], it is shown that there
always exists a sufficiently small sampling period
T,
for which
the deviations between the actual
PWM
controlled responses
(of an integral equation) and those of an averaged model,
under identical initial conditions, remain arbitrarily close to
each other. This,
ol;
course, is an immediate consequence of
continuity on
T.
Therefore, it seems reasonable to approach the problem
of approximating the dynamics of (2.1) by using classical
averaging techniques on integral equation (2.2). Classical aver-
aging techniques have the advantage over the techniques
of
[4]
the validity of the approximation. By performing averaging
on an integral equation instead of a differential equation, this
paper will show that the difficulties due to many types of state
discontinuities are eliminated. This approach allows a rigorous
explanation, which was not provided in [3] and [9].
Most classical averaging techniques
[6],
[7], though, are not
directly applicable to integral equations. However, recently,
new state space averaging theory has been developed that
relies entirely on the representation of solutions of differential
equations by their corresponding integral equation [11], [12].
The results of
[
1
11
and
[
121 are written for infinite dimensional
dynamical systems, but the techniques, as this paper shows,
can also be applied to ordinary differential equations.
111. AVERAGING
OF STATE
DISCONTINUOUS
POWER
ELECTRONIC
SYSTEMS
In general form, feedback pulse width modulated systems
considered in this paper will be modeled by the integral
equation
"( t;
t o,
x(t o))
=
Z ( t )
.
u( dL( z( s) )
-
tri
(s,
T ) )
]
ds
(3.1)
where it will alwaiys be assumed that
z
E
R",
t o
denotes
initial time, and
f,:
R"
+
R"
are locally Lipschitz functions,
i.e., there exists an open neighborhood
R
c
R"
such that for
every
z1
E
R,
x2
E
R,
there are constant positive
k,
satisfying
l/fi(.l)
-f,(z2)11
5
kz l/z l
-z21/.
The functions
d,:
Rn
-+
R
are the duty ratios and will also be assumed locally Lipschitz in
R
with Lipschitz constant
m,.
Furthermore, they will always
satisfy
0
5
dL( z)
5
1.
Along with
(3. I
),
consider the corresponding "averaged"
integral equation
y(t;
t o,
Y(t 0))
=y(t )
544
IEEE
TRANSACTIONS ON
POWER
ELECTRONICS,
VOL.
11,
NO.
4,
JULY
1996
where
f t,
and
d,
are
as
previously defined and
y
E
R".
This
section will discuss the conditions under which solutions to
(3.2) can approximate solutions to (3.1). Since (3.2) is both
continuous and autonomous, its analysis is much simpler than
that of discontinuous and nonautonomous (3.1). For example,
if
f L
and
d,
have continuous partial derivatives with respect to
z,
then the stability properties of (3.2) may be determined
by examining the eigenvalues of the linearization of (3.2)
No
such simple statement can be said
about determining the stability of (3.1). The two theorems
presented in this section extend the results of [3] to the state
discontinuous case, i.e., to the feedback control case.
A.
Chattering
By representing state discontinuous differential equations by
a
corresponding integral equation, it is possible to rigorously
explain averaging approximations in power electronic systems.
However, it will always be necessary to assume that the
models under consideration have a finite number of right-
hand side state discontinuities on any bounded time interval
and that each discontinuity is Lebesgue integrable. This,
however, is not always true for mathematical models of power
electronic systems. For example, when systems are switching
infinitely often (chattering), there exists no compact time
interval in which the right-hand side of the state discontinuous
differential equation is continuous. Hence, a unique solution
to a corresponding integral equation will not exist in the usual
sense unless the theory
of
differential inclusions
[SI
is used.
In
this paper, we will always assume that the system is
not chattering. The physical implication of this assumption is
that power electronic switches turn on and off only once each
PWM
switching period. Conditions for guaranteeing this are
presented in [13] and will not be discussed here. However, it
is important
to
note that the averaging results presented below,
are only valid when chattering does not occur.
B.
Theoretical Results
We begin this section
by
outlining the general averaging
procedure that will be taken in this paper to justify the
approximation of (3.1) by (3.2).
Given
a
nonautonomous, integral equation [such as (2.1) or
(3.1)]
z ( t )
=
z(to)
+
Ji
g( s,
~ ( s ),
T )
ds,
consider the cor-
responding autonomous "averaged' integral equation
y(t)
=
y(t0)
+
JL
g(y(s))ds,
where
g(.)
is an "average value" of
y ( t,
.,
.)
and
J (.)
does not depend on time,
t,
or on the
switching period,
T.
Step 1: Take
the difference between the two integral equa-
tions to obtain
Step
2:
Show that for any
6
>
0,
however small, and any
L
>
t o,
however large, there will always exist a
To
=
To(6,
L)
and
a
constant
K
>
0
such that for
0
<
T
5
TO
for any
t
E
[ t o, L].
Step
3:
Immediately from Step
1,
Step 2, and Gronwall's
inequality, this implies that for
t
E
[ t o,
L],
L
>
t o,
and
0 < T s T o
l l 4t)
-
Y( t ) l l
I
(Ilz(t0)
-
y(to)ll
+
6) eK( L- t o)
where
6
i
0
as
T
4
0.
This implies that on any arbitrarily
large but bounded time interval, if
z(t0)
=
y(to),
then
x( t )
and
y(t)
can remain arbitrarily close to each other for a sufficiently
small switching period.
Step
4:
Assume that
z(to)
= y( t 0)
and that
y(t)
ap-
proaches a uniformly asymptotically stable equilibrium point,
y3.
Then, there will always exist
a
sufficiently small
To
=
To(6)
such that, for
0
<
T
5
TO
lIz(t)
-
Y( t )/l
<
6,
t
2
t o.
Furthermore, this result will remain valid for initial conditions
that satisfy
l I ~ ( t 0 )
-
y( t o) l l
I
p,
where
p
>
0
is sufficiently
small.
Step
4
basically states that if averaging can be proven
on a finite time interval, then it can always be extended
to an infinite time interval in the special case when the
averaged solution approaches
a
uniformly asymptotically sta-
ble equilibrium point. This statement
has
been proven by
many authors [3],
[6],
171,
[12] and is standard to averaging
theory.
Once Step
2
is completed, Steps 3 and
4
will immediately
follow. However, it turns out that, for
PWM
systems, com-
pleting Step 2
is
extremely difficult and relies
on
some very
recently developed mathematical tools
[
1 11,
[
121. Keeping the
above algorithm in mind, it is now possible to prove the
main results
of
this paper. The proof of Theorem 3.1 relies
on several Lemmatta, which are presented in the Appendix.
Theorem 3.1: Let
z(t )
and
y(C)
denote the solutions to (3.1)
and (3.2), respectively. Then, for any constant
L
>
t o
and for
any constant
7
>
0,
there exists
a
TO
=
To(q,
L)
>
0
and a
constant
K
>
0
such that, for
0
<
T
5
To,
ll4t)
- Y( t ) l l
I
(Ilz(t0)
-y(to)ll+rl)
exp{K(t-to)}
(3.3)
for all
t
E
[ t o,
L].
3:
R"
--f
R"
and
W:
R"
-+
R"
as
Proof
of Theorem
3.1:
For simplicity, define operators
.
u( di ( z( s) )
-
tri
(s,
T ) )
ds
(3.4)
LEHMAN
AND
BASS:
EXTENSIONS
OF AVERAGING
THEORY
FOR POWER ELECTRONIC
SYSTEMS
Under the assumption of no chattering,
x( t ),
the solution to
(3.1) will be continuous. Therefore, it
is
well known (Theo-
rems 24.4 and
24.5,
[14]) that
x( t )
can be approximated by
piecewise constant functions. Construct
N
+
1
such piecewise
constant functions
&(t)
E
R",
z
=
0,
1, .
.. ,
N,
such that
for
any
t
E
[t o,
L],
0
5
[ d,( z ( t ) )
-
dZ( &( t ) ) ]
5
S,
for
i
=
1,
2,
. .
.
,
N,
where
6,
>
0
are a set of positive constants.
Furthermore, choose
&(t)
such that for any
t
E
[ t o,
L], Il f L(x(t ))
-
f,(Za(t))lI
I
6,
also, for
i
=
0,
1, 2,
. .
.
,
N.
Since
f i (.)
and
rip(.)
are Lipschitz
functions, such
&(t)
can always be constructed for arbitrary
6,
>
0.
Define
( J Z ) ( t )
as
(J?)(t)
=z(to)
+
f o(&(s))
ds
lot
u(di(Zi(s))
-
tri
(s,
T ) )
ds.
Consider
(3.6)
-
u( di ( Zi ( s) )
-
tri
( s, T))ll
ds.
(3.7)
By Lemma A.2, for any
t
E
[t o,
L],
Ilfi(&(t))lJ
5
Mi;
i
=
1,
2,
...
,
N.
Let
M
=
max{Mi};
i
=
1, 2,
...
,
N.
Then, using the fact that
11u(.)11
_<
1
and using the fact that
Zi
have been constructed
so
that
Ilfi(x(t))
-
fi(Zi(t))ll
_<
Si
for any
t
E
[ t o,
L],
i
=
0,
1, 2,
.
. .
,
N,
(3.7) becomes
u( d;( &( s) ) -
tri
( s,
T))ll
ds
(3.8)
for any
t
E
[ t o,
L].
However,
&(.)
and
6,
have been cho-
sen
so
that
d,(Z.,(t))
5
d,(x(t ))
5
d,( &( t ) )
+
S,,
i
=
I,
2, ...
,
N
for any
t.
Define
N
new piecewise constant
functions,
h%(Z-i (t )),
where
h,(&(t))
=
min(1,
d,(Z,(t))+
S,};
i
=
1,
2,
...
,
N.
Note that
d,(&(t ))
5
d,( z ( t ) )
5
~
545
h,(lc,(t))
for all
t
ci
( t o,
L);
i
=
1,
2,
...
,
N.
Then, by
Lemma
A. 1,
this implies
II(J-x)(t)
-
(J-lc)(t>ll
I
So(N
+
l)(t
-
t o )
N t
+
M
E;
lo
I I U(ha(&(S))
-
tri(s,
TI)
,=I
(3.9)
Using Lemma
AS.
there will always exist a
TO
=
To(a,, L)
-
u(d,(i.,(s))
-
tri
(s,
T))ll
ds.
such that, for
0
<
'I'
5
TO
Il(J-x)(t)
-
( J4( t ) I l
5
So(N
+
1)(t
-
t o )
N
+
M
[a,
+
S,(t
-
t o ) ]
2 =1
=
0
+
Yl(6)
(3.10)
where
cr
=
M
E,=]
cr,
goes to zero as
T
+
0,
S
=
[ SO,
.
. .
,
6 ~ ]
and
y1(S)
is a positive constant that approaches
zero as
6,
4
0.
1v
Similarly, for any
t
E
[ t o,
L]
Noting that
I l f,( &( t ) ) l l
I
M,
for any
t
E
[t o,
L]
and that
lld,(.)11
5
1,
(3.11)
becomes
N
II(W4(t)
-
(Wn:)(t)lI
5
So(t
-
-
t o )
2=1
N
+
&,(t
-
t o )
2 = 1
I
Yz
(6)
,
for any
t
E
[ t o,
L].
Clearly
yz(6)
4
0
as
6,
+
0.
Consider now the inequality
II(Jx)(t)
-
(Wx)(t)II
5
Il(Jz>(t)
-
( W( t ) l I
+
II(JZ)(t)
-
(wz)(t)lI
+
II(W?)(t)
-
(Wx)(t)II
(3.12)
which is true for all
t.
Using the above discussion and
Lemma
A.4,
there
exists
a
TO
=
To(cr,
0,
L)
such that, for
~
z( t )
is
continuous. from basic averaging theory, it can be derived that
TO
is
546
0
<
T
5
To,
where
CJ
and
p
are positive constants that approach zero
as
T
-+
0,
as defined in (3.10) and Lemma
A.4,
respectively.
Constants
yl(6)
and
r~(6)
can be made arbitrary small by
making
&(.)
approximate
z(
.)
with arbitrary accuracy. There-
fore, without loss of generality, it can be assumed that
6,
+
0,
which implies that for a sufficiently small switching period
where
=
c7
+
p,
and
+
0
as
T
+
0.
Finally, consider the inequality
The following is always true:
Noting that
f;
(.)
and
di
(.)
are Lipschitz and that
0
5
d;
(.)
I
1,
one obtains
where
m,
are the Lipschitz constants of
d,(.)
and
k,
are
the Lipschitz constants for
f,(.).
Let
K
=
M
m,
+
E,"=,
IC,.
Then (3.15) becomes
for any
t
E
[ t o,
L].
Applying Gronwall's inequality completes
the proof of the theorem. Q.E.D.
Remark
3.1:
The main trick of the proof of Theorem 3.1
is
to construct
N
+
1
piecewise constant functions
fii(t),
i
=
0,
1, ...
,
N,
which accurately approximate
z ( t )
on
t
E
[ t o,
L].
Such functions can always be constructed since
IEEE TRANSACTIONS
ON
POWER ELECTRONICS,
VOL.
11,
NO.
4,
JULY
1996
Then, using the notation defined in (3.4)-(3.6) and (3.1
1)
Now, Step
2
of the averaging algorithm must be performed.
Each term on the right-hand side of (3.17) is considered
separately. By constructing
&( t )
to approximate
z( t )
with
arbitrary accuracy, the quantities
I I
(Jz)
(t )
-
( 3 2 )
(t )
I I
and
II(W?)(t)
-
(Wx)(t)I/
can be made arbitrarily small. In
essence, this is due to the Fundamental Theorem of Calculus,
which states that any integral can be estimated by the sums
of the areas of rectangles. Since
&(t)
is piecewise constant,
(J?)(t )
and
(W?)(t)
represent nothing more than areas
under the curve of a piecewise constant function which is
equivalent to summing the areas of rectangles. Of course,
due to the discontinuities that appear in
J(.),
theoretically arguments must be made in order to justify these
approximations.
Likewise, because
f,
(.)
and
d,
(.)
have been assumed Lips-
chitz, it is not too difficult to show that for any
t
E
[ t o,
L]
Now, the only term left to consider in (3.17) is
II(J%)(t)
-
(W2)
(t )
/
1.
However, this term only considers the difference
between the integrals of piecewise constant functions, which,
as the theorem shows, is a much simpler problem to handle
(based on the lemmatta in the Appendix).
Remark
3.2:
When
.(to)
= y( t a),
Theorem 3.1 guarantees
that there will always exist a sufficiently small switching
period such that for any
17
>
0,
however small,
I I z( t ) - y( t ) l
I
<
7
on any finite time interval. This bound
is
true, even when
(3.1) or (3.2) are unstable. For the case when solutions are
bounded, however, more powerful theorems can be stated.
Remark
3.3:
The choice
of
TO
is best found through numer-
ical simulation, since theoretical estimates are often extremely
conservative. One reason for poor theoretical estimates
of
TO
is that Theorem
3.1
does not distinguish between stable and
unstable systems. For unstable systems, it is possible that
solutions to (3.1) and (3.2) grow exponentially, making it
difficult to estimate the difference,
l l x(t )
-
y(t)ll. With this
in mind, we make these general statements:
For
general systems, from the proof
of
Theorem
3.1
and
LEHMAN AND
BASS:
EXTENSIONS
OF AVERAGING
THEORY
FOR
POWER ELECTRONIC SYSTEMS
541
sufficiently small if all three of the following conditions are
satisfied:
1) there exists no chattering in the system;
2)
TO
<<
e--kc(Lpt o),
where
k,
are the Lipschitz constants
3)
To
<<
where
m,
are the Lipschitz constants
This is not to say that for every system in question, the
switching period must be chosen
so
that 1)-3) are satisfied.
For example, if solutions to (3.2) decay exponentially to an
equilibrium point, then condition 2) can often be relaxed. It is
important to remark that condition
1)
must always be fulfilled
or else the solutions
of
(3.1) will not be defined in the usual
sense.
Remark3.4:
Based on the Theorem 3.1 and the above
discussion, it is possible to determine general conditions that
suggest the improvement of the accuracy of approximation
between the original (3.1) and the approximate (3.2) system.
Clearly, the approximation becomes better
as
the switching
period becomes smaller, but also,
as
Remark 3.3 notes, the
approximations will tend to improve for systems with smaller
Lipschitz constants, i.e., the smaller
k,
and
m,
are, the
more accurate the averaging technique will tend to be (for
general systems) and the better for linear systems than for
nonlinear systems. Additionally, as Theorem 3.2 suggests
below, if the averaged system is stable, then the averaging
approximations will also improve. Conversely, if the averaged
system is unstable, the averaging approximation tends to
worsen. Finally,
as
is clear from (3.3), a necessary condition
for the solutions of (3.2) to approximate the solutions of (3.3)
is that the initial conditions of the two systems must be chosen
in appropriate neighborhoods.
Remark
3.5:
One of the main advantages of the averaging
technique is that nonlinearities are maintained in the averaged
system. Hence, the approximation of (3.1) by (3.2) is valid
even when the states,
x,
become large, which would not be
true if
a
linearization technique were to be used. The averaging
approximation is, therefore, valid for large signals.
As
stated earlier, when the solution to the averaged equa-
tion approaches a uniformly asymptotically stable equilibrium
point, the solutions of (3.1) and of (3.2) will remain close
to each other on an infinite time interval for a sufficiently
small switching period. The following theorem is an immediate
consequence of this fact. The proof is almost identical to
Proposition
4
of [3] or Theorem 2.2 of [12], and therefore,
is omitted.
Theorem
3.2:
Let
~ ( t )
and
y ( t )
denote the solutions to (3.1)
and (3.2), respectively, and let
ys
E
R
(ys
#
y ( t 0) )
denote
a
uniformly asymptotically stable equilibrium point. Suppose
that
y(t)
i
ys
as
t
4
00.
Then there are constants
Po(q)
and
To(q)
such that, for any
for
f,(.);
for
d,(.).
v
>
0,
any
I l 4 t o )
-
Y( t 0) l l
<
P,
0
5
P
<
Po
<
v>
and
O < T < T o
Il4t)
-
Y( t ) l l
<
r/
(3.19)
for all
t 2
t o.
Remark
3.6:
The above theorem gives conditions in which
the interval
in
Theorem 3.1 can be made infinite. For the
case when
y ( t )
apprloaches a uniformly asymptotically stable
equilibrium point,
y..,
the difference,
Ilz(t)
-
y( t ) l l,
can be
t
2
t o
assuming
I
Iz(t0)
-
y(to)
I I
and the switching period are sufficiently small.
Remark
3.7:
Suppose
f,(.)
and
di (.)
have continuous par-
tial derivatives. Then, for an equilibrium point,
ys,
of (3.2) to
be uniformly asymptotically stable, it is possible to check that
8.f
0
(Ys
Det
SI
-
--
{
ay
have all solutions with Re(s)
<
0.
Remark
3.8:
Theorem 3.2 guarantees that under the proper
conditions, when (3.2) is stable, then
so
is (3.1). Unlike
(3.2), however, the solution
to
(3.1) will not in general
approach an equilibrium point
as
t
+
00,
since (3.1) is
a
time varying integral equation. In general, the solution to
(3.1) will (assuming it is stable) approach
a
periodic orbit.
However, this periodic orbit will not necessarily be in the
vicinity of the equilibrium point of the averaged equation,
unless
T
is sufficiently small. In fact, (the theory clearly
shows that) it is possible to construct examples in which
(3.1) has an asymptotically stable periodic orbit for
all
T,
but is only in the vicinity of the equilibrium point of (3.2)
when
T
+
0
(see Section IV). This behavior becomes more
pronounced in feedback controlled
(as
opposed to open loop)
PWM dc-dc converters due to the nonlinearities, and is not
noted in [3] and
[9].
We further explain this phenomenon in
[lo].
Remark 3.9:
In Theorems 3.1 and 3.2, the feedback signals
are compared with tri(t,
T),
shown in Fig.
1.
However, all
the above theorems remain valid for triangle waves
as
shown
in Fig. 2
also,
provided that they are rescaled to vary between
zero and one (see Section IV). Furthermore, it is not necessary
to compare each
d,
(.)
with the same function with the same pe-
riod. For instance, in (3.1) we might have
U( &(.)
-
tri
(.,
T,))
U( &(.)
-
tri
(.,
T) ),
where
T,
might not equal
T3,
for
i
#
j.
As
long
as
each
T,
is sufficiently small, all previous
results will remain valid.
C.
Ripple Estimate
It is often desirable to obtain an estimate on the ripple of
the system, which will be denoted in this paper as
Q( t,
T,
.).
Then, practical applications of averaging tell us that a better
approximation of the solution to
(3.1)
will be given by
where
~ ( t )
and
y(t)
are the solutions of (3.1) and
(3.2),
respectively,
T
is thie switching period, and
Q(t,
T, .)
is
the
ripple estimate obtained by the following algorithm.
548
IEEE TRANSACTIONS
ON
POWER ELECTRONICS,
VOL.
11,
NO.
4,
JULY
1996
(4
Fig. 2. Other possible triangle waves.
Consider only the right-hand sides
of
(3.1) and (3.2). Let
z(t0)
=
y( t o),
and replace every
z(s)
and
y(s)
in (3.1) and
(3.2) by the constant
c
E
72". Now take the difference between
(3.1)
and (3.2)
to
obtain
N
(3.21)
where
s,
h(t )
dt
denotes the indefinite integral of
h(t )
(math-
ematically referred to
as
the primitive). The ripple estimate is
given
as
1
r 7
Replacing
c
by
y ( t )
yields
Q(t,
T,
.).
Performing integrations
(3.21) and (3.22), using (3.1) and (3.2), an estimate on the
ripple is computed to be
As
the switching period becomes smaller, the amplitude
of
q(t,
T,
.)
will also become smaller and ripple of the
on the initial condition can be made by solving the equation
.(to)
= y( t o)
+
Q( t,
T,
&t o ) ),
for
y(t0)
in terms
of
.(to).
The general expression for the ripple estimate (3.23) is an
important contribution
of
this work and has been used in [lo]
to help model the effects of switching at lower frequencies.
IV. APPLICATION EXAMPLE
Consider the
PWM
boost converter with feedback control
structure
as
shown in Fig. 3. Open-loop operation
of
this con-
verter was considered in
[3]:
however, the theory developed
in
[3]
does not extend to closed-loop operation
(as
do the
theorems in this paper). Assuming the converter is operating
in the continuous conduction mode, the closed loop (rescaled)
system description is given by
E
b =
[f]
where the components of
x( t )
=
[ i ~ ( t ), vc(t)lT
are
the
inductor current and capacitor voltage. Note,
that
since the
triangle wave in Fig.
3
varies from 0.7-3
V,
it
is
necessary to
rescale the system into (4.1)
so
that Theorems 3.1 and 3.2 can
be applied. This is easily done scaling the duty ratio function
using the minimum (trimin
=
0.7
V)
and maximum (trimax
=
3.0
V)
values
of
the triangle wave:
g(z)
-
trimin
trimax
-
trimin
d(x)
=
where
g(z)
is defined in Fig.
3.
For this specific
g(z),
we have
VREF
=
0.312.3,
kl
=
0.412.3, and
=
-0.112.3.
LEHMAN
AND
BASS:
EXTENSIONS
OF
AVERAGING THEORY
FOR
POWER ELECTRONIC
SYSTEMS
E= 5V
L=
50p.H
i&)
I I
g(x>
=
1
-0.4
iL(t)
+
0.1
vc(t)
-0.4
0.7
+
2.3
tri
(t,
Fig. 3. Feedback
control
boost
converter
Application of Theorems 3.1 and 3.2 is now immediate upon
noting that, using the previous notation,
fo(z)
=
Aoz
+
b,
f l ( x )
=
Alz,
and
N
=
1.
The closed loop switching and
averaged models were simulated using Saber [15]. Fig. 4 illus-
trates the switching and averaged trajectories of the capacitor
voltage for different switching periods. As the frequency of the
system,
f,?,
increases, or equivalently as the switching period
decreases (since
f s
=
T1),
the approximation of
z(t )
by
y ( t )
improves. For example, when
f s
=
50
KHz,
system (4.1) has
proximately) 7.3
V.
The averaged system, on the other hand,
approaches (approximately)
8.5
V. As
the frequency of the
system increases (the switching period decreases) the capacitor
voltage for (4.2) more closely approximates the capacitor
voltage of (4.1). For
f s
= 1
MHz,
state capacitor voltage that oscillates about (approximately)
8.4
V,
representing
a
for larger frequency, the amplitude of the ripple decreases.
This further verifies Theorems 3.1 and
3.2,
which state that the
approximation between the averaged system and the original
system improves as the switching period decreases and is
consistent with Remark 3.8. Similar results can be obtained
for the inductor current.
549
I
Using (3.23), it is possible to directly compute an estimate
on the ripple
of
the system as
Fig.
5
plots the capacitor voltage and inductor current of the
original system (4.1) when
f s
=
100
KHz.
A comparison of
these plots can be m,ade with Fig. 6, which
shows
the improve-
ment of the averaging technique by approximating
z(t )
by
x( t )
E
y ( t )
+
@(t,
T,
y( t ) )
and updating the initial condition,
y( t o),
by solving [given
.(to)]
the nonlinear equation
Fig. 6 indicates that the “shape”
of
solutions to averaged
to
the ripple estimate closely resembles the
“shape” of solutionrs to the original system (plus, perhaps, a dc
offset). Therefore, ,the ripple estimate may provide important
system information, even at a low frequency (large switching
period).
550
OU
t ( S )
( V)
9-
8.5 -
8.
7.5.
7.
6.5.
6.
5.5.
5.
4.5.
1-
900111-
IEEE
TRANSACTIONS
ON
POWER
ELECTRONICS. VOL.
11,
NO. 4,
JULY
1996
9-
8
5 -
4
100m-
0-
I
24,
56,
7:~
l Obu
12\u
1 5'0 ~
17:u
20011
22\11
25bu 275u
3
I
4
5 -
4
Fig. 4.
_ - _ -
fs
=
100
kHz;
- - - -
f.
=
50
kHz;
- - - - -
average.
Simulated start up transient response of capacitor voltage for (4.1) and (4.2) for different values of switching frequency.
__
fa
=
1
MHz;
111
t ( s )
Fig. 5. Simulated start up transient response of both capacitor voltage and inductor current for
(4.1)
when switching frequency equals
100
kHz.
__
vc(t);
- - - -
iZ(t).
V.
CONCLUSION
A
rigorous averaging theory for power electronic systems
has been developed. This new theory extends previous work
to
include state discontinuous (feedback controlled)
PWM
systems. The two theorems Presented in this Paper Provide
a
basis
for
LEHMAN
AND BASS:
EXTENSIONS
OF
AVERAGING THEORY FOR
POWER
ELECTRONIC
SYSTEMS
551
1-
9.
900m-
8.5.
800111-
8-
7oom-
7.5-
600111-
7-
500111-
6.5 -
400111-
6.
300m-
5.5 -
Zoom-
5.
100m-
4.5.
0-
4-
d
Fig.
6.
equals
100
KHz.
__
vcavg(t);
- - - -
vc
r1p
(t ).
,
- -
- -
-
alaVg(t);
-
-
- -
-
- drl p(t );.
Simulated
start
up
transient response
of both
capacitor voltage and inductor
current
for
(4.2)
and
ripple
correction
(4.3) when
switching frequency
model and its relation to the original switching model. First-
order ripple estimates are derived, and an application
of
the
theory to a feedback controlled boost converter is presented.
APPENDIX
LemmaA.1:
Let
g1(x)
and
gz(x)
be functions mapping
Then, for any
x
E
Rn,
any
T
>
0,
and any
t
E
R
the
2"
-+
R.
Suppose that, for any
2
E
R",
SI(\$)
5
g2( 5).
following inequality
is
always true:
.(91(4
-
tri
(t,
T) )
I
.(g2(4
-
tri
(4
TI).
(-4.1)
Proof
of
Lemma
A.1:
If
g1
(x)
_<
gZ(z),
then at no time
can
gI(s)
-
hi
(t,
T)
>
0
while
g2(x)
-
tri
( t,
T )
<
0.
Using
this
fact and applying the definition of the Heaviside step
function, the proof is immediate.
Q.E.D.
Lemma
A.2:
Suppose that
z(t)
and
y( t )
are given
by
(3.1)
and
(3.2),
respectively. Then for any
t E
[ t o,
L], L
2
t o
Il x(t)l l
L
I l ~ ( t 0 ) l l
exp
Ilv(t)ll
L
I l Y(t 0)l l exp
where k,
are
the Lipschitz constants
of
fi,
previously defined.
ProofofLemma
A.2:
By
(3.1
j
-t
5
.I'
Ilfi(+>)ll
2=1
t o
.
l l ~( dz ( ~( s ) )
-
tri
( s, T))ll
ds.
(A.4)
Since
ft
are Lipschitz functions with Lipschitz constants k,
and since
llu(-)\l
5
1,
we have
which, by Gronwall's inequality, implies
(A.2).
Upon noting
that
lldz(.)ll
5
I,
(A.3)
can be obtained using almost the same
arguments.
Q.E.D.
LemmaA.3: Let
.D
be a constant satisfying
0
5
D
5
1.
Then,
for
any
t
2
t o
111;
[u(D
-
tri
(s.
T) )
-
D]
ds
5
( ( DT( 1
-
D)ll.
(A.6)
Proof
of
Lemma
A.3: Without
loss
of
generality, assume
that
t o
=
0 (initial time can
always
be
redefined so that this
is the case.)
By
definition
II
1
0
t E
[nT,
nT
+
DT]
t E
[aT
+
DT,
(n
f
1)TI
%(I?
-
tri(t, T) )
=
(A.7)
n
=
0,
1,
2,
... .
Assume that D
f
0.
(The
case when D
=
0
is trivially
proved since both the left and right-hand side
of
(A.6) are
identically equal to zero). Suppose
0
5
t
_<
DT.
Then
552
IEEE TRANSACTIONS ON
POWER
ELECTRONICS,
VOL.
11,
NO.
4,
JULY
1996
=
I
l(1
-
D)tl
I
I
IIDT(1
-
o)ll,
( ~,8 )
Proof
of
Lemma
A.4:
If
c
is a constant vector, then
d;(c)
and
f z( c)
are constants also. Therefore
By Lemma A.3, there exists a
To =
To(yi,
L)
such that,
(A.9)
for
o
<
T
5
To
Finally, suppose that
t
2
T.
Then, there always exists an
N
integer
M
=
M( t,
T),
depending on
t
and
T,
such that
M
2
1
and
MT
5
t
5
( M
+
1)T. Therefore
I/
[u(D
-
tri
( s,
T) )
-
D]
ds
=
111""
[u(D
-
tri
(s,
7'))
-
D]
ds
[u(D
-
tri(s,
5"))
-
D]
ds
.
(A.lO)
I 1
t
+
/MT
MT
Due to periodicity,
so
[u( D
-
tri
( s,
T))
-
D]
ds
=
M
s:[u(D-tri(s, T) ) - D] ds andJ hT [u(D-tri(s,
7'))-
D] ds
=
J,
[u( D
-
tri
(s,
T) )
-
D]
ds. Note that
t-MT
where
M(.)
is defined in (A.15), and
T~
are arbitrary
small
positive constants. From here, it follows that
N
i=l
for any
t o
5
t l
5
t 2
5
L.
Since
z(t )
is a piecewise constant function, there will
always exist a sequence
to
=
ao
<
a1
<
a2
<
...
<
up
=
t,
t
I
L,
and a set
of
constants
{ c 3};
j
=
1, 2,
...
,
p,
with
cI
=
5(t )
on the interval
t E
[u3-1,
a3],
such that
which completes the proof. Q.E.D. (A. 19)
Lemma
A.4:
Let
5(t )
be a piecewise constant function.
Then for any constant
L
>
to
and any constant
p
>
0,
there
Noting that
SUP,
IIfz(c~)1I
5
SUP,
Ilc311
<
00,
it is easy
exists
a
7'0
=
To(P,
L)
such that, for
0
<
T
5
TO
to see that (A.19) can be made arbitrarily small by making
7%
arbitrarily small (by choosing
To
sufficiently small). Defining
"g
l;
fr(z(s))'LL(dz(z(s))
-
t'i
( s,
TI) ds
r = l
l l
N
P
=
f
( P
+
1)
SUP
Ilf2(c3>llrz
z = 1
3
f z ( ~ ( s ) ) d r ( z ( s ) )
ds
5
P;
t
E
[to,
LI.
J
=
1, 2,
...
,p
(A.20)
(A.
14)
the proof is complete.
Q.E.D.
LEHMAN AND BASS: EXTENSIONS
OF
AVERAGING
THEORY
FOR POWER ELECTRONIC
SYSTEMS
553
Lemma
A.5:
Let g1
(z)
and
92( 2)
be continuous functions
mapping
Rn
-+
R,
with
0
5
gI(z)
5
1
and
0
5
g2(2)
5
1.
Suppose that
5(t )
is a piecewise constant function and that
0
5
gl ( s( t >)
-
g2(5(t))
5
6,
for
some
constant
6
>
0 and
for all
t
E
[t o,
L], L
>
t o.
Then for any constant
L
>
t o
and any constant
0
>
0,
there
exists a constant
To
=
To(a,
L)
such that, for
0
<
T
5
TO
Proof of
Lemma
A.5:
By Lemma A.
1
and simple algebra
(A.23)
By Lemma
A.4,
for any
y
>
0,
there exist a
TO
=
T0( y,
L)
such that,
for
0
<
T
5
To,
i
=
1,
2.
(A.24)
Defining
0
=
2y
and noting that
11g1(2(t))
-
g2(2(t))ll
5
S
for all
t
E
[t o,
L],
(A.23) immediately gives (A.21).
Q.E.D.
REFERENCES
R.
D.
Middlebrook and
S.
Cuk, “A general unified approach
to
mod-
eling switching-converter power stages,” in ZEEE Power Electronics
Specialists Con\$ R,ec., 1976, pp. 18-34.
B.
Y.
Lau and R. D. Middlebrook, “Small-signal frequency response
theory for piecewise-constant two switched-network dc-to-dc converter
systems,” in IEEE Power Specialists
Con\$
Rec., 1986, pp. 186-200.
P. T. Krein,
J.
Bentsman,
R.
M. Bass, and
B.
C.
Lesieutre, “On the use
of averaging for the analysis of power electronic systems,” IEEE Trans.
Power Electron., vsol. 5, pp. 182-190, Apr. 1990.
H.
Sira Ramirez, “A geometric approach to pulse-width modulated
systems,” ZEEE Truns. Autom. Contr., vol. 34,
no.
2,
pp. 184-187, Feb.
1989.
D. G.
Taylor,
“Pulse
width modulated control
of
electromechanical
systems,”
IEEE
Trans. Automat. Contr., vol. 37,
no.
4,
pp. 524-528,
Apr. 1992.
N. N. Bogoliubov and
Y.
A. Mitropolsky, Asymptotic Methods in the
Theory
of Nonlinear Oscillations. New York Gordon and Breach,
1961.
S.
M. Meerkov, “Averaging of trajectories of slow dynamic systems,”
DifSerential Equdons, vol. 9, no. 11, pp. 1239-1245, 1973.
A.
F.
Filippov, Differential Equations with Discontinuous Righthand
Sides.
The
J. Sun
and
H.
Grotstollen, “Averaged modeling and analysis of resonant
converters,” in IE,EE Power Electronics Specialists Con\$ Rec., 1993,
B.
Lehman and R. M.
Bass,
“Switching frequency dependent averaged
models for
PWM
dc-dc converters,” ZEEE Trans. Power Electron., vol.
11, pp. 89-98,
Jan.
1996.
B. Lehman, “Averaging
of
differential delay equations,” in Proc.
1992
ZEEE ACC,
June
1992, vol. 2, pp. 1955-1956.
B. Lehman,
J.
Bentsman,
S.
Verduyn Lunel, and E. Vemest, “Vibrational
control of nonlinear time lag systems: Averaging theory, stabilizability,
and transient response,” IEEE Trans. Automat. Contr., vol. 39,
no.
5,
pp. 898-912, May 1994.
R.
M.
Bass and
P.
‘T.
Krein, “Large-signal design alternatives for switch-
ing power converter control,” in IEEE Power Electronics Specialists
Con\$
Rec., June 1991, pp. 882-887.
R. G.
Bartle, The Elements
of
Real Analysis. New York: Wiley, 1976.
Introduction to Saber Simulator, Analogy, 1993.
pp. 707-713.
for
a
photograph and biography,
see
p. 98 of the January
1996 issue of this
TRANSACTIONS.
Richard
M.
Bass
(S’82-M’82-SM’94),
for a photograph and biography,
see
p.
98 of
the
January 1996 issue of
this
TRANSACTIONS.